Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Approximate Model Predictive Control for Microgrid Energy Management via Imitation Learning

Changrui Liu, Shengling Shi, Anil Alan, Ganesh Kumar Venayagamoorthy, Bart De Schutter

TL;DR

This work addresses real-time microgrid energy management under uncertainty by directly learning a policy that imitates mixed-integer EMPC. It proposes a novel imitation-learning framework that uses features capturing ESS state, previous generator outputs, and forecast-induced disturbances, along with noise injection to mitigate distribution shift. The method trains an MLP to map augmented state information to current control inputs, followed by projection to satisfy input constraints, achieving comparable economic performance to EMPC while reducing online computation to about 10% of solving the MIQP. The approach is demonstrated on a PV/WT and two-fuel-generator microgrid, highlighting practical viability for fast, robust real-time EMS in high-renewable settings.

Abstract

Efficient energy management is essential for reliable and sustainable microgrid operation amid increasing renewable integration. This paper proposes an imitation learning-based framework to approximate mixed-integer Economic Model Predictive Control (EMPC) for microgrid energy management. The proposed method trains a neural network to imitate expert EMPC control actions from offline trajectories, enabling fast, real-time decision making without solving optimization problems online. To enhance robustness and generalization, the learning process includes noise injection during training to mitigate distribution shift and explicitly incorporates forecast uncertainty in renewable generation and demand. Simulation results demonstrate that the learned policy achieves economic performance comparable to EMPC while only requiring $10\%$ of the computation time of optimization-based EMPC in practice.

Approximate Model Predictive Control for Microgrid Energy Management via Imitation Learning

TL;DR

This work addresses real-time microgrid energy management under uncertainty by directly learning a policy that imitates mixed-integer EMPC. It proposes a novel imitation-learning framework that uses features capturing ESS state, previous generator outputs, and forecast-induced disturbances, along with noise injection to mitigate distribution shift. The method trains an MLP to map augmented state information to current control inputs, followed by projection to satisfy input constraints, achieving comparable economic performance to EMPC while reducing online computation to about 10% of solving the MIQP. The approach is demonstrated on a PV/WT and two-fuel-generator microgrid, highlighting practical viability for fast, robust real-time EMS in high-renewable settings.

Abstract

Efficient energy management is essential for reliable and sustainable microgrid operation amid increasing renewable integration. This paper proposes an imitation learning-based framework to approximate mixed-integer Economic Model Predictive Control (EMPC) for microgrid energy management. The proposed method trains a neural network to imitate expert EMPC control actions from offline trajectories, enabling fast, real-time decision making without solving optimization problems online. To enhance robustness and generalization, the learning process includes noise injection during training to mitigate distribution shift and explicitly incorporates forecast uncertainty in renewable generation and demand. Simulation results demonstrate that the learned policy achieves economic performance comparable to EMPC while only requiring of the computation time of optimization-based EMPC in practice.
Paper Structure (16 sections, 38 equations, 6 figures)

This paper contains 16 sections, 38 equations, 6 figures.

Table of Contents

  1. Introduction
  2. Microgrid Modeling
  3. Fuel Generators
  4. Energy Storage System
  5. Renewable Energy Sources
  6. Curtailable Loads
  7. Power Exchange & Power Balance
  8. Economic Model Predictive Control for Microgrid Energy Management
  9. Direct Approximate Mixed-Integer Model Predictive Control via Imitation Learning
  10. Motivations
  11. Imitation Learning for Approximate EMPC
  12. Feature Design:
  13. Data Generation With Noise Injection:
  14. Loss Function, Training & Post Processing:
  15. Case Study
  16. ...and 1 more sections

Figures (6)

  • Figure 1: The considered microgrid configuration illustrating the control actions (blue arrows) executed by the Energy Management System (EMS): fuel generator output, load curtailment, and energy exchange with the main grid.
  • Figure 2: Schematic overview of the fully connected MLP with some representative weight links, where the red neurons with output value $1$ are used to model biases.
  • Figure 3: Visualization of $50$ disturbance realizations showing the power generated by photovoltaic panels (PV) and wind turbines (WT), as well as the power consumed by the load. The dark dotted line represents the nominal disturbance realization adopted from gao2021online, while the shaded area indicates the range between the maximum and minimum values.
  • Figure 4: Nominal constant $c_{\mathrm{p}}$ and $c_{\mathrm{s}}$ with their noisy time-varying profiles for simulation.
  • Figure 5: Input comparison of (a) the optimization-based expert EMPC (EMPC), (b) the proposed IL-based approximate EMPC (Proposed IL), and (c) the baseline IL-based approximate EMPC (Baseline IL).
  • ...and 1 more figures